Using multidimensional histogram equalization as relative radiometric calibration for change detection in remote sensing imagery

摘要

Change detection in remote sensing provides useful information for various applications. The purpose is to locate the changed area between two remotely sensed images collected at different times and identify the materials before and after the changes. However, the difference in spectrum may not solely result from the changes of the endmembers on the ground. The spectrum of the same material in two remote sensing images may not be the same due to the different condition of solar illumination and atmosphere condition while the images were obtained. Therefore, radiometric calibration is required before applying the change detection algorithm and comparing the spectrum. Since it is difficult to have complete information for absolute radiometric calibration, in this paper, we propose a relative radiometric calibration method to transform the data, so these images from different times will have similar radiation and atmosphere effect.The histogram specification is used to adjust the gray level distribution of one image to match that of another. The process is performed by equalizing the histogram and followed by inversion of the cumulated histogram. Since multispectral imagery has multiple dimensions, its histogram is also multidimensional. The conventional algorithm assumes each component is independent, so it performs 1-D histogram equalization to each component separately. But the assumption is usually not true, the spectrum is correlated between bands in multispectral images. In this study, we perform the histogram equalization in a multidimensional space. Since a given histogram can be viewed as a Gaussian mixture, in order to fit the uniform distribution, each index of the histogram needed to be relocated. The difference of the original histogram and the uniform distribution is demanded to be minimum. The results of the histogram equalization can be used to accomplish histogram specification between the two images.Finally, we adopt two SPOT-5 images for our experiment, and we also compare the experimental results with previous proposed "decorrelation stretch."